An evaluation of feed-forward, back-propagation neural networks for optimal design of structural systems and for modular analysis of electronic packages

Artificial neural network is well known for its ability to capture arbitrarily complex non-linear functions. As one of the most popular network structures, feed-forward, back-propagation network has been used extensively in solving general engineering design/analysis problems and will be studied in this thesis work. In particular, The use of neural networks in solving structural optimization problems and a general engineering analysis problem found in electronic packaging is explored in detail.; The first half of this thesis, from Chapter 1 to 3, will be focusing on providing a systematic comparison of the efficiency and accuracy of the neural network-based solution schemes to classical structural optimization techniques. An extensive literature review on structural optimization using neural networks is presented in Chapter 1. In Chapter 2, the neural network training procedures used in the present evaluation are described in detail. When using first-order nonlinear programming algorithms with neural networks, the ability to approximate derivatives is important. Therefore, mainly for completeness of evaluation, two new training methods that use the derivative information are proposed in addition to the now common function-based training method. The first method uses the derivatives to create additional training points in the vicinity of the original points, based on Taylor's series expansion. The second method attempts to minimize the error in derivatives while imposing the error in output functions as constraint. Expressions for analytical derivatives are derived for both function-based and derivative-based training. Significant savings in computational time are reported when calculating derivatives using built-in analytical derivatives instead of using finite difference derivatives. In Chapter 3 the proposed methods are applied to solve five optimization problems with varying degree of complexity. Approximately 1100 test cases are executed to compare the accuracy and efficiency of neural network-based optimization with the classical approaches. With careful training, the neural networks were found capable of predicting the exact solution, but this ability depended considerably on the complexity of the original problem. In general, the accuracy of the network, which is a function of the training effort, suffered considerably in large problems. (Abstract shortened by UMI.)...